| | |
| | | #include "convolutional_layer.h" |
| | | #include "utils.h" |
| | | #include "mini_blas.h" |
| | | #include "im2col.h" |
| | | #include "col2im.h" |
| | | #include "blas.h" |
| | | #include "gemm.h" |
| | | #include <stdio.h> |
| | | #include <time.h> |
| | | |
| | |
| | | return float_to_image(h,w,c,layer.delta); |
| | | } |
| | | |
| | | convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, float learning_rate, float momentum, float decay) |
| | | convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation) |
| | | { |
| | | int i; |
| | | size = 2*(size/2)+1; //HA! And you thought you'd use an even sized filter... |
| | | convolutional_layer *layer = calloc(1, sizeof(convolutional_layer)); |
| | | |
| | | layer->learning_rate = learning_rate; |
| | | layer->momentum = momentum; |
| | | layer->decay = decay; |
| | | |
| | | layer->h = h; |
| | | layer->w = w; |
| | | layer->c = c; |
| | |
| | | |
| | | layer->filters = calloc(c*n*size*size, sizeof(float)); |
| | | layer->filter_updates = calloc(c*n*size*size, sizeof(float)); |
| | | layer->filter_momentum = calloc(c*n*size*size, sizeof(float)); |
| | | |
| | | layer->biases = calloc(n, sizeof(float)); |
| | | layer->bias_updates = calloc(n, sizeof(float)); |
| | | layer->bias_momentum = calloc(n, sizeof(float)); |
| | | float scale = 1./(size*size*c); |
| | | scale = .05; |
| | | for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*2*(rand_uniform()-.5); |
| | | float scale = 1./sqrt(size*size*c); |
| | | for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*rand_normal(); |
| | | for(i = 0; i < n; ++i){ |
| | | //layer->biases[i] = rand_normal()*scale + scale; |
| | | layer->biases[i] = .5; |
| | | layer->biases[i] = scale; |
| | | } |
| | | int out_h = convolutional_out_height(*layer); |
| | | int out_w = convolutional_out_width(*layer); |
| | | |
| | | layer->col_image = calloc(layer->batch*out_h*out_w*size*size*c, sizeof(float)); |
| | | layer->col_image = calloc(out_h*out_w*size*size*c, sizeof(float)); |
| | | layer->output = calloc(layer->batch*out_h * out_w * n, sizeof(float)); |
| | | layer->delta = calloc(layer->batch*out_h * out_w * n, sizeof(float)); |
| | | |
| | | #ifdef GPU |
| | | layer->filters_cl = cl_make_array(layer->filters, c*n*size*size); |
| | | layer->filter_updates_cl = cl_make_array(layer->filter_updates, c*n*size*size); |
| | | layer->filter_momentum_cl = cl_make_array(layer->filter_momentum, c*n*size*size); |
| | | layer->filters_gpu = cuda_make_array(layer->filters, c*n*size*size); |
| | | layer->filter_updates_gpu = cuda_make_array(layer->filter_updates, c*n*size*size); |
| | | |
| | | layer->biases_cl = cl_make_array(layer->biases, n); |
| | | layer->bias_updates_cl = cl_make_array(layer->bias_updates, n); |
| | | layer->bias_momentum_cl = cl_make_array(layer->bias_momentum, n); |
| | | layer->biases_gpu = cuda_make_array(layer->biases, n); |
| | | layer->bias_updates_gpu = cuda_make_array(layer->bias_updates, n); |
| | | |
| | | layer->col_image_cl = cl_make_array(layer->col_image, layer->batch*out_h*out_w*size*size*c); |
| | | layer->delta_cl = cl_make_array(layer->delta, layer->batch*out_h*out_w*n); |
| | | layer->output_cl = cl_make_array(layer->output, layer->batch*out_h*out_w*n); |
| | | layer->col_image_gpu = cuda_make_array(layer->col_image, out_h*out_w*size*size*c); |
| | | layer->delta_gpu = cuda_make_array(layer->delta, layer->batch*out_h*out_w*n); |
| | | layer->output_gpu = cuda_make_array(layer->output, layer->batch*out_h*out_w*n); |
| | | #endif |
| | | layer->activation = activation; |
| | | |
| | |
| | | return layer; |
| | | } |
| | | |
| | | void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c) |
| | | void resize_convolutional_layer(convolutional_layer *layer, int h, int w) |
| | | { |
| | | layer->h = h; |
| | | layer->w = w; |
| | | layer->c = c; |
| | | int out_h = convolutional_out_height(*layer); |
| | | int out_w = convolutional_out_width(*layer); |
| | | |
| | | layer->col_image = realloc(layer->col_image, |
| | | layer->batch*out_h*out_w*layer->size*layer->size*layer->c*sizeof(float)); |
| | | out_h*out_w*layer->size*layer->size*layer->c*sizeof(float)); |
| | | layer->output = realloc(layer->output, |
| | | layer->batch*out_h * out_w * layer->n*sizeof(float)); |
| | | layer->delta = realloc(layer->delta, |
| | | layer->batch*out_h * out_w * layer->n*sizeof(float)); |
| | | |
| | | #ifdef GPU |
| | | cuda_free(layer->col_image_gpu); |
| | | cuda_free(layer->delta_gpu); |
| | | cuda_free(layer->output_gpu); |
| | | |
| | | layer->col_image_gpu = cuda_make_array(layer->col_image, out_h*out_w*layer->size*layer->size*layer->c); |
| | | layer->delta_gpu = cuda_make_array(layer->delta, layer->batch*out_h*out_w*layer->n); |
| | | layer->output_gpu = cuda_make_array(layer->output, layer->batch*out_h*out_w*layer->n); |
| | | #endif |
| | | } |
| | | |
| | | void bias_output(const convolutional_layer layer) |
| | | void bias_output(float *output, float *biases, int batch, int n, int size) |
| | | { |
| | | int i,j,b; |
| | | int out_h = convolutional_out_height(layer); |
| | | int out_w = convolutional_out_width(layer); |
| | | for(b = 0; b < layer.batch; ++b){ |
| | | for(i = 0; i < layer.n; ++i){ |
| | | for(j = 0; j < out_h*out_w; ++j){ |
| | | layer.output[(b*layer.n + i)*out_h*out_w + j] = layer.biases[i]; |
| | | for(b = 0; b < batch; ++b){ |
| | | for(i = 0; i < n; ++i){ |
| | | for(j = 0; j < size; ++j){ |
| | | output[(b*n + i)*size + j] = biases[i]; |
| | | } |
| | | } |
| | | } |
| | | } |
| | | |
| | | void forward_convolutional_layer(const convolutional_layer layer, float *in) |
| | | void backward_bias(float *bias_updates, float *delta, int batch, int n, int size) |
| | | { |
| | | float alpha = 1./batch; |
| | | int i,b; |
| | | for(b = 0; b < batch; ++b){ |
| | | for(i = 0; i < n; ++i){ |
| | | bias_updates[i] += alpha * sum_array(delta+size*(i+b*n), size); |
| | | } |
| | | } |
| | | } |
| | | |
| | | |
| | | void forward_convolutional_layer(const convolutional_layer layer, network_state state) |
| | | { |
| | | int out_h = convolutional_out_height(layer); |
| | | int out_w = convolutional_out_width(layer); |
| | | int i; |
| | | |
| | | bias_output(layer); |
| | | bias_output(layer.output, layer.biases, layer.batch, layer.n, out_h*out_w); |
| | | |
| | | int m = layer.n; |
| | | int k = layer.size*layer.size*layer.c; |
| | |
| | | float *b = layer.col_image; |
| | | float *c = layer.output; |
| | | |
| | | im2col_cpu(in, layer.batch, layer.c, layer.h, layer.w, |
| | | layer.size, layer.stride, layer.pad, b); |
| | | |
| | | for(i = 0; i < layer.batch; ++i){ |
| | | im2col_cpu(state.input, layer.c, layer.h, layer.w, |
| | | layer.size, layer.stride, layer.pad, b); |
| | | gemm(0,0,m,n,k,1,a,k,b,n,1,c,n); |
| | | b += k*n; |
| | | c += n*m; |
| | | state.input += layer.c*layer.h*layer.w; |
| | | } |
| | | activate_array(layer.output, m*n*layer.batch, layer.activation); |
| | | } |
| | | |
| | | void learn_bias_convolutional_layer(convolutional_layer layer) |
| | | void backward_convolutional_layer(convolutional_layer layer, network_state state) |
| | | { |
| | | int i,b; |
| | | int size = convolutional_out_height(layer) |
| | | *convolutional_out_width(layer); |
| | | for(b = 0; b < layer.batch; ++b){ |
| | | for(i = 0; i < layer.n; ++i){ |
| | | layer.bias_updates[i] += sum_array(layer.delta+size*(i+b*layer.n), size); |
| | | } |
| | | } |
| | | } |
| | | |
| | | void backward_convolutional_layer(convolutional_layer layer, float *delta) |
| | | { |
| | | float alpha = 1./layer.batch; |
| | | int i; |
| | | int m = layer.n; |
| | | int n = layer.size*layer.size*layer.c; |
| | | int k = convolutional_out_height(layer)* |
| | | convolutional_out_width(layer); |
| | | gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta); |
| | | learn_bias_convolutional_layer(layer); |
| | | |
| | | float *a = layer.delta; |
| | | float *b = layer.col_image; |
| | | float *c = layer.filter_updates; |
| | | gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta); |
| | | backward_bias(layer.bias_updates, layer.delta, layer.batch, layer.n, k); |
| | | |
| | | if(state.delta) memset(state.delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float)); |
| | | |
| | | for(i = 0; i < layer.batch; ++i){ |
| | | gemm(0,1,m,n,k,1,a,k,b,k,1,c,n); |
| | | a += m*k; |
| | | b += k*n; |
| | | } |
| | | float *a = layer.delta + i*m*k; |
| | | float *b = layer.col_image; |
| | | float *c = layer.filter_updates; |
| | | |
| | | if(delta){ |
| | | m = layer.size*layer.size*layer.c; |
| | | k = layer.n; |
| | | n = convolutional_out_height(layer)* |
| | | convolutional_out_width(layer); |
| | | float *im = state.input+i*layer.c*layer.h*layer.w; |
| | | |
| | | a = layer.filters; |
| | | b = layer.delta; |
| | | c = layer.col_image; |
| | | im2col_cpu(im, layer.c, layer.h, layer.w, |
| | | layer.size, layer.stride, layer.pad, b); |
| | | gemm(0,1,m,n,k,alpha,a,k,b,k,1,c,n); |
| | | |
| | | for(i = 0; i < layer.batch; ++i){ |
| | | gemm(1,0,m,n,k,1,a,m,b,n,0,c,n); |
| | | b += k*n; |
| | | c += m*n; |
| | | if(state.delta){ |
| | | a = layer.filters; |
| | | b = layer.delta + i*m*k; |
| | | c = layer.col_image; |
| | | |
| | | gemm(1,0,n,k,m,1,a,n,b,k,0,c,k); |
| | | |
| | | col2im_cpu(layer.col_image, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, state.delta+i*layer.c*layer.h*layer.w); |
| | | } |
| | | |
| | | memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float)); |
| | | |
| | | col2im_cpu(layer.col_image, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta); |
| | | } |
| | | } |
| | | |
| | | void update_convolutional_layer(convolutional_layer layer) |
| | | void update_convolutional_layer(convolutional_layer layer, float learning_rate, float momentum, float decay) |
| | | { |
| | | int size = layer.size*layer.size*layer.c*layer.n; |
| | | axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1); |
| | | scal_cpu(layer.n, layer.momentum, layer.bias_updates, 1); |
| | | axpy_cpu(layer.n, learning_rate, layer.bias_updates, 1, layer.biases, 1); |
| | | scal_cpu(layer.n, momentum, layer.bias_updates, 1); |
| | | |
| | | scal_cpu(size, 1.-layer.learning_rate*layer.decay, layer.filters, 1); |
| | | axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1); |
| | | scal_cpu(size, layer.momentum, layer.filter_updates, 1); |
| | | axpy_cpu(size, -decay, layer.filters, 1, layer.filter_updates, 1); |
| | | axpy_cpu(size, learning_rate, layer.filter_updates, 1, layer.filters, 1); |
| | | scal_cpu(size, momentum, layer.filter_updates, 1); |
| | | } |
| | | |
| | | |
| | |
| | | return single_filters; |
| | | } |
| | | |
| | | #ifdef GPU |
| | | |
| | | cl_kernel get_convolutional_learn_bias_kernel() |
| | | { |
| | | static int init = 0; |
| | | static cl_kernel kernel; |
| | | if(!init){ |
| | | kernel = get_kernel("src/convolutional_layer.cl", "learn_bias", 0); |
| | | init = 1; |
| | | } |
| | | return kernel; |
| | | } |
| | | |
| | | void learn_bias_convolutional_layer_ongpu(convolutional_layer layer) |
| | | { |
| | | int size = convolutional_out_height(layer) * convolutional_out_width(layer); |
| | | |
| | | cl_setup(); |
| | | cl_kernel kernel = get_convolutional_learn_bias_kernel(); |
| | | cl_command_queue queue = cl.queue; |
| | | |
| | | cl_uint i = 0; |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(layer.batch), (void*) &layer.batch); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(layer.n), (void*) &layer.n); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(size), (void*) &size); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(layer.delta_cl), (void*) &layer.delta_cl); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(layer.bias_updates_cl), (void*) &layer.bias_updates_cl); |
| | | check_error(cl); |
| | | |
| | | const size_t global_size[] = {layer.n}; |
| | | |
| | | clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0); |
| | | check_error(cl); |
| | | } |
| | | |
| | | cl_kernel get_convolutional_bias_kernel() |
| | | { |
| | | static int init = 0; |
| | | static cl_kernel kernel; |
| | | if(!init){ |
| | | kernel = get_kernel("src/convolutional_layer.cl", "bias", 0); |
| | | init = 1; |
| | | } |
| | | return kernel; |
| | | } |
| | | |
| | | void bias_output_gpu(const convolutional_layer layer) |
| | | { |
| | | int out_h = convolutional_out_height(layer); |
| | | int out_w = convolutional_out_width(layer); |
| | | int size = out_h*out_w; |
| | | |
| | | cl_setup(); |
| | | cl_kernel kernel = get_convolutional_bias_kernel(); |
| | | cl_command_queue queue = cl.queue; |
| | | |
| | | cl_uint i = 0; |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(layer.n), (void*) &layer.n); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(size), (void*) &size); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(layer.biases_cl), (void*) &layer.biases_cl); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(layer.output_cl), (void*) &layer.output_cl); |
| | | check_error(cl); |
| | | |
| | | const size_t global_size[] = {layer.batch, layer.n*size}; |
| | | |
| | | clEnqueueNDRangeKernel(queue, kernel, 2, 0, global_size, 0, 0, 0, 0); |
| | | check_error(cl); |
| | | } |
| | | |
| | | //#define TIMEIT |
| | | |
| | | void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in) |
| | | { |
| | | int i; |
| | | int m = layer.n; |
| | | int k = layer.size*layer.size*layer.c; |
| | | int n = convolutional_out_height(layer)* |
| | | convolutional_out_width(layer); |
| | | |
| | | bias_output_gpu(layer); |
| | | |
| | | #ifdef TIMEIT |
| | | clock_t time = clock(); |
| | | printf("Forward\n"); |
| | | #endif |
| | | |
| | | im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl); |
| | | |
| | | #ifdef TIMEIT |
| | | clFinish(cl.queue); |
| | | printf("Im2col %f\n", sec(clock()-time)); |
| | | time = clock(); |
| | | #endif |
| | | |
| | | for(i = 0; i < layer.batch; ++i){ |
| | | cl_mem a = layer.filters_cl; |
| | | cl_mem b = layer.col_image_cl; |
| | | cl_mem c = layer.output_cl; |
| | | gemm_ongpu_offset(0,0,m,n,k,1.,a,0,k,b,i*k*n,n,1.,c,i*m*n,n); |
| | | } |
| | | #ifdef TIMEIT |
| | | clFinish(cl.queue); |
| | | printf("Gemm %f\n", sec(clock()-time)); |
| | | #endif |
| | | activate_array_ongpu(layer.output_cl, m*n*layer.batch, layer.activation); |
| | | #ifdef TIMEIT |
| | | cl_read_array(layer.output_cl, layer.output, m*n*layer.batch); |
| | | #endif |
| | | } |
| | | |
| | | void backward_convolutional_layer_gpu(convolutional_layer layer, cl_mem delta_cl) |
| | | { |
| | | int i; |
| | | int m = layer.n; |
| | | int n = layer.size*layer.size*layer.c; |
| | | int k = convolutional_out_height(layer)* |
| | | convolutional_out_width(layer); |
| | | gradient_array_ongpu(layer.output_cl, m*k*layer.batch, layer.activation, layer.delta_cl); |
| | | learn_bias_convolutional_layer_ongpu(layer); |
| | | |
| | | for(i = 0; i < layer.batch; ++i){ |
| | | cl_mem a = layer.delta_cl; |
| | | cl_mem b = layer.col_image_cl; |
| | | cl_mem c = layer.filter_updates_cl; |
| | | |
| | | gemm_ongpu_offset(0,1,m,n,k,1,a,i*m*k,k,b,i*k*n,k,1,c,0,n); |
| | | } |
| | | //cl_read_array(layer.delta_cl, layer.delta, m*k*layer.batch); |
| | | |
| | | if(delta_cl){ |
| | | m = layer.size*layer.size*layer.c; |
| | | k = layer.n; |
| | | n = convolutional_out_height(layer)* |
| | | convolutional_out_width(layer); |
| | | |
| | | for(i = 0; i < layer.batch; ++i){ |
| | | cl_mem a = layer.filters_cl; |
| | | cl_mem b = layer.delta_cl; |
| | | cl_mem c = layer.col_image_cl; |
| | | |
| | | gemm_ongpu_offset(1,0,m,n,k,1,a,0,m,b,i*k*n,n,0,c,i*m*n,n); |
| | | } |
| | | |
| | | scal_ongpu(layer.batch*layer.h*layer.w*layer.c,0,delta_cl, 1); |
| | | col2im_ongpu(layer.col_image_cl, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_cl); |
| | | } |
| | | } |
| | | |
| | | void pull_convolutional_layer(convolutional_layer layer) |
| | | { |
| | | cl_read_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size); |
| | | cl_read_array(layer.biases_cl, layer.biases, layer.n); |
| | | } |
| | | |
| | | void push_convolutional_layer(convolutional_layer layer) |
| | | { |
| | | cl_write_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size); |
| | | cl_write_array(layer.biases_cl, layer.biases, layer.n); |
| | | } |
| | | |
| | | void update_convolutional_layer_gpu(convolutional_layer layer) |
| | | { |
| | | int size = layer.size*layer.size*layer.c*layer.n; |
| | | axpy_ongpu(layer.n, layer.learning_rate, layer.bias_updates_cl, 1, layer.biases_cl, 1); |
| | | scal_ongpu(layer.n,layer.momentum, layer.bias_updates_cl, 1); |
| | | |
| | | scal_ongpu(size, 1.-layer.learning_rate*layer.decay, layer.filters_cl, 1); |
| | | axpy_ongpu(size, layer.learning_rate, layer.filter_updates_cl, 1, layer.filters_cl, 1); |
| | | scal_ongpu(size, layer.momentum, layer.filter_updates_cl, 1); |
| | | pull_convolutional_layer(layer); |
| | | } |
| | | |
| | | |
| | | #endif |
| | | |